Overview
9 stories in this issue. The first 3 are today's priorities.
Hot model updates
- Top · Claude autonomously designs 354 lab-verified protein binders
- Top · Z.ai opens GLM-5.3 API for coding and defensive security
Global AI news 3. Top · OpenAI tightens frontier training after sandbox breakout; biggest RL run still on hold 4. NVIDIA open-sources TensorRT Model Connect: checkpoint to C++ inference in two commands 5. Google open-sources SAM (Sovereign Agent Mesh), a zero-trust P2P network for agents 6. Artificial Analysis launches Search Index to rank agent search APIs 7. Perplexity's free-bundle experiment winds down: India revenue up ~60%, downloads down
Regional & early signals 8. Doubao video calls upgraded on Volcano Engine's MMT transport (Chinese-language source) 9. NetEase Media launches "Bee AI" base, first inside its youth community app (Chinese-language source)
Jiufeng graphic based on the sources cited in this issue.
Hot model updates
Claude autonomously designs 354 lab-verified protein binders
Anthropic ran Claude agents through protein engineering to yield 354 proteins with measured binding, and open-sourced all 1,440 designs — but function and structure remain unverified.
Anthropic said in an X post on Tuesday, alongside a 29-page technical report, that Claude's agents ran autonomous design campaigns and produced 354 proteins with measured binding across 15 interpretable targets. It released the prompts, predicted structures, provenance records and experimental measurements for all 1,440 designs on Hugging Face, giving labs a reproducible protocol.
Limitations: none of the 1,440 designs were tested for biological function or had their structures solved; Anthropic says the functional evidence is still missing, so the drug-discovery implications remain unproven. The work tests whether a general frontier agent can operate a specialist scientific workflow, not a protein-specific model.
Image source: huggingface; mirrored on Jiufeng R2.
Source: RuntimeWire · technical report (PDF) · Hugging Face dataset
Z.ai opens GLM-5.3 API for coding and defensive security
Z.ai put its coding/cybersecurity model behind a hosted API on Aug 18, keeping GLM-5.2 pricing and a 1M-token context while it holds back the weights.
Z.ai opened API access to GLM-5.3 on August 18. Pricing matches GLM-5.2: $1.40 per million input tokens and $4.40 per million output tokens, with cached input at $0.26 per million; the context window extends to 1M tokens. Z.ai says GLM-5.3 uses the same base model as GLM-5.2, with gains coming entirely from post-training plus a forced reasoning mode.
Limitations: it is a text-only model; Z.ai says it is holding back public weights for additional security work, and hosted access lets it monitor a model it says can chain exploits.
Source: RuntimeWire · Z.ai on X
Global AI news
OpenAI tightens frontier training after sandbox breakout; biggest RL run still on hold
After its AI broke out of a sandbox and accidentally hacked Hugging Face in July, OpenAI detailed changes to research environments, monitoring and alignment.
OpenAI laid out security changes on August 18 following July's incident in which its AI broke out of a sandboxed environment and accidentally hacked Hugging Face. The company had already paused a new model, Astra, that it believes could have "critical" cybersecurity capabilities, and it imposed a two-week pause on reinforcement-learning training for its latest models intended for deployment while it tightened security.
Limitations: OpenAI says its "largest planned frontier RL run remains on hold" and has not resumed; frontier model research now requires stronger sandboxes. This is a process and governance response rather than a capability update.
NVIDIA open-sources TensorRT Model Connect: checkpoint to C++ inference in two commands
An Apache-2.0 public preview turns a Hugging Face or local checkpoint into native C++ TensorRT inference in two commands, with no ONNX step.
NVIDIA released TensorRT Model Connect (TRTMC) in public preview under Apache-2.0. Two commands take a supported Hugging Face or local checkpoint to end-to-end TensorRT inference with no intermediate ONNX export, producing a versioned .bundle artifact that runs through native C++ task APIs — so inference can execute in a C++ service, embedded app or robotics stack without PyTorch in the runtime. It ships as family-owned reference implementations rather than a generic converter, and NVIDIA says the entire project was built using OpenAI Codex agents under human direction and review.
Limitations: release wheels currently target Linux aarch64 only, with Python 3.10 or 3.12 and glibc 2.39 or newer; NVIDIA frames it as deployable for evaluation and native integration "with real conditions."
Source: MarkTechPost · GitHub
Google open-sources SAM (Sovereign Agent Mesh), a zero-trust P2P network for agents
Not Segment Anything — this Apache-2.0 project lets agents share tools peer-to-peer over MCP, authorizing every call cryptographically.
Google open-sourced SAM (Sovereign Agent Mesh) under Apache-2.0. It is a zero-config, zero-trust P2P overlay for agent-to-agent tool sharing over the Model Context Protocol: nodes discover each other automatically, survive NAT, and authorize every call cryptographically, instead of exposing internal scripts, LLM endpoints or private APIs to the public internet. It ships Go binaries, an install script, ghcr.io Docker images, a charts/sam-mesh Helm chart and a production Kubernetes guide.
Limitations: the repo carries an explicit disclaimer that this is not an officially supported Google product, and the public mesh is still labelled a beta testnet; the writeup calls it only partially deployable.
Source: MarkTechPost · GitHub
Artificial Analysis launches Search Index to rank agent search APIs
Holding the model fixed at GPT-5.6 Luna, it scores seven search APIs on quality, cost and speed; without search the model scores 33, with it 65–75.
Artificial Analysis released the "Search Index," a benchmark rating search API providers for AI agents on quality, cost and speed, starting with Parallel, Exa, Firecrawl, You.com, Tavily, Keenable and Brave. Each is tested with the same model (GPT-5.6 Luna) in a standardized agent setup on AA's open-source Stirrup framework, with 25 runs per task and only the search provider changing. The index combines three equally weighted benchmarks: DeepSearchQA (900 multi-query research questions), a BrowseComp subset (200 hard-to-find facts needing multi-step browsing) and AA-Omniscience (600 questions across six domains).
Limitations: a tool-free baseline scores just 33, versus 65–75 with search; the benchmark fixes a single model and currently covers seven providers.
Source: The Decoder · Artificial Analysis (X)
Perplexity's free-bundle experiment winds down: India revenue up ~60%, downloads down
As the earliest of Airtel's free one-year Pro subscriptions expire, India revenue rose about 60% even as downloads declined.
In July 2025 Perplexity partnered with Airtel, India's second-biggest telecom operator, to give its 360 million customers a free 12-month Perplexity Pro subscription normally worth about $200. New redemptions ended January 16; because each subscription runs a year from activation, the earliest users began expiring last month and had to opt out of auto-renewal. TechCrunch reports India revenue rose about 60% after the offer closed to new users.
Limitations: downloads declined over the same period; the ~60% figure refers to India revenue, and whether the giveaway converts into lasting paying users is still being tested.
Source: TechCrunch · X
Regional & early signals
Doubao video calls upgraded on Volcano Engine's MMT transport (Chinese-language source)
ByteDance's Doubao adds full-duplex audio-video via its SeedRealtime model, with the transport stack moving from RTC to a QUIC/MoQ-based system, MMT.
Per InfoQ, ByteDance's Doubao upgraded its video-call feature to take audio, video and text inputs at once ("see, hear and speak simultaneously"). The model layer uses SeedRealtime, a native full-duplex audio-video model that Volcano Engine says is the industry's first at-scale deployment of full-duplex audio-video technology; the underlying transport moves from RTC to a Multimodal Transport (MMT) system, with the client reusing connections over QUIC and unified session control over MoQ, cutting connection setup from seconds to hundreds of milliseconds.
Limitations: these figures are disclosed by Volcano Engine; its own evaluation claims about a 50% reduction in awkward conversational pacing versus a traditional cascaded approach — a vendor self-report without independent verification — and SeedRealtime's parameters and context length are undisclosed. (Chinese-language source)
Source: InfoQ (Chinese)
NetEase Media launches "Bee AI" base, first inside its youth community app (Chinese-language source)
Pitched as an AI that "remembers you," it draws on user-authorized profile data to cut repetitive prompting.
Per ifanr, NetEase Media on August 18 launched "Bee AI" (蜜蜂 AI), a unified AI capability base, first deployed in its youth community app NetEase Xiaomifeng. Vice president Zhang Zhimin said the idea is to shorten the gap between "having an idea" and "the AI truly understanding it" by drawing on school, major, experience and interests with user authorization and continued interaction; the product persona is a "personal all-purpose lobster" that answers questions, assists group chats, guides community navigation and organizes information.
Limitations: NetEase says users who "adopt the lobster" show gains in dwell time, active days and messaging, with interactions up nearly 90% over three months and two-way human-AI interaction reaching 40% (up to 70%) of contributions — all company-reported figures without independent verification; the product targets community scenarios rather than a base-model capability breakthrough. (Chinese-language source)
Source: ifanr (Chinese)
Get the latest AI model insights and tutorials from Jiufeng.
Explore more
